Are Time Series Foundation Models Ready for Vital Sign Forecasting in Healthcare?

Xiao Gu, Yu Liu, Zaineb Mohsin, Jonathan Bedford, Anshul Thakur, Peter Watkinson, Lei Clifton, Tingting Zhu, David Clifton
Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:401-419, 2025.

Abstract

The rise of foundation models, particularly large language models like ChatGPT, has revolutionized natural language processing and demonstrated remarkable generalization across numerous healthcare applications. Building on this success, foundation models for time series forecasting have emerged, offering new opportunities by leveraging pretraining on large-scale datasets. However, existing time series foundation models are pretrained with minimal clinical data, and their potentials for continuously recorded clinical time series, such as vital signs, remain largely under-explored. This motivates our endeavor to integrate time series foundation models with vital sign data to address critical clinical challenges, particularly in predicting patient deterioration. Through an extensive evaluation of various settings and configurations of these models, alongside comparisons with conventional forecasting models, we highlight the significant opportunities for improvement in developing clinically useful time series forecasting models. In a word, the “ChatGPT” moment for time series foundation models, in the typical clinical domain, is yet to come.

Cite this Paper


BibTeX
@InProceedings{pmlr-v259-gu25a, title = {Are Time Series Foundation Models Ready for Vital Sign Forecasting in Healthcare?}, author = {Gu, Xiao and Liu, Yu and Mohsin, Zaineb and Bedford, Jonathan and Thakur, Anshul and Watkinson, Peter and Clifton, Lei and Zhu, Tingting and Clifton, David}, booktitle = {Proceedings of the 4th Machine Learning for Health Symposium}, pages = {401--419}, year = {2025}, editor = {Hegselmann, Stefan and Zhou, Helen and Healey, Elizabeth and Chang, Trenton and Ellington, Caleb and Mhasawade, Vishwali and Tonekaboni, Sana and Argaw, Peniel and Zhang, Haoran}, volume = {259}, series = {Proceedings of Machine Learning Research}, month = {15--16 Dec}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v259/main/assets/gu25a/gu25a.pdf}, url = {https://proceedings.mlr.press/v259/gu25a.html}, abstract = {The rise of foundation models, particularly large language models like ChatGPT, has revolutionized natural language processing and demonstrated remarkable generalization across numerous healthcare applications. Building on this success, foundation models for time series forecasting have emerged, offering new opportunities by leveraging pretraining on large-scale datasets. However, existing time series foundation models are pretrained with minimal clinical data, and their potentials for continuously recorded clinical time series, such as vital signs, remain largely under-explored. This motivates our endeavor to integrate time series foundation models with vital sign data to address critical clinical challenges, particularly in predicting patient deterioration. Through an extensive evaluation of various settings and configurations of these models, alongside comparisons with conventional forecasting models, we highlight the significant opportunities for improvement in developing clinically useful time series forecasting models. In a word, the “ChatGPT” moment for time series foundation models, in the typical clinical domain, is yet to come.} }
Endnote
%0 Conference Paper %T Are Time Series Foundation Models Ready for Vital Sign Forecasting in Healthcare? %A Xiao Gu %A Yu Liu %A Zaineb Mohsin %A Jonathan Bedford %A Anshul Thakur %A Peter Watkinson %A Lei Clifton %A Tingting Zhu %A David Clifton %B Proceedings of the 4th Machine Learning for Health Symposium %C Proceedings of Machine Learning Research %D 2025 %E Stefan Hegselmann %E Helen Zhou %E Elizabeth Healey %E Trenton Chang %E Caleb Ellington %E Vishwali Mhasawade %E Sana Tonekaboni %E Peniel Argaw %E Haoran Zhang %F pmlr-v259-gu25a %I PMLR %P 401--419 %U https://proceedings.mlr.press/v259/gu25a.html %V 259 %X The rise of foundation models, particularly large language models like ChatGPT, has revolutionized natural language processing and demonstrated remarkable generalization across numerous healthcare applications. Building on this success, foundation models for time series forecasting have emerged, offering new opportunities by leveraging pretraining on large-scale datasets. However, existing time series foundation models are pretrained with minimal clinical data, and their potentials for continuously recorded clinical time series, such as vital signs, remain largely under-explored. This motivates our endeavor to integrate time series foundation models with vital sign data to address critical clinical challenges, particularly in predicting patient deterioration. Through an extensive evaluation of various settings and configurations of these models, alongside comparisons with conventional forecasting models, we highlight the significant opportunities for improvement in developing clinically useful time series forecasting models. In a word, the “ChatGPT” moment for time series foundation models, in the typical clinical domain, is yet to come.
APA
Gu, X., Liu, Y., Mohsin, Z., Bedford, J., Thakur, A., Watkinson, P., Clifton, L., Zhu, T. & Clifton, D.. (2025). Are Time Series Foundation Models Ready for Vital Sign Forecasting in Healthcare?. Proceedings of the 4th Machine Learning for Health Symposium, in Proceedings of Machine Learning Research 259:401-419 Available from https://proceedings.mlr.press/v259/gu25a.html.

Related Material